Peter Gjertsson, Lena Johansson, Milan Lomsky, Mattias Ohlsson, S. Richard Underwood and Lars Edenbrandt
Clinical data do not improve artificial neural network interpretation of myocardial perfusion scintigraphy
LU TP 09-07
Abstract:
Purpose: Artificial neural networks interpretation of myocardial perfusion scintigraphy (MPS) has so far been based on image data alone. Physicians reporting MPS often combine image and clinical data. The aim was to evaluate if neural network interpretation would be improved by adding clinical data to image data.
Methods: 418 patients were used for training and 532 patients for testing the neural networks. First the network was trained with image data alone and thereafter with image data in combination with clinical parameters (age, gender, previous infarction, coronary artery by-pass grafting, percutaneous coronary intervention, typical chest pain, present smoker, hypertension, hyperlipedemia, diabetes, peripheral vascular disease, positive family history). Expert interpretation was used as gold standard. Receiver operating characteristic (ROC) curves were calculated and the areas under the ROC curves for the networks trained with and without clinical data were compared for the diagnosis of myocardial infarction and ischaemia.
Results: There was no statistically significant difference in ROC area for the diagnosis of myocardial infarction between the neural network trained with the combination of clinical and image data (95.2%) and with image data alone (96.5%). For the diagnosis of ischaemia there was no statistically significant difference in ROC area between the neural network trained with the combination of clinical and image data (83.2.3%) and with image data alone (84.4%).
Conclusion: Neural network interpretation of MPS is not improved when clinical data is added to perfusion and functional data. One reason for this could be that experts base their interpretations of MPS mainly on the images and to a lesser degree on clinical data.
Purpose: Artificial neural networks interpretation of myocardial perfusion scintigraphy (MPS) has so far been based on image data alone. Physicians reporting MPS often combine image and clinical data. The aim was to evaluate if neural network interpretation would be improved by adding clinical data to image data.
Methods: 418 patients were used for training and 532 patients for testing the neural networks. First the network was trained with image data alone and thereafter with image data in combination with clinical parameters (age, gender, previous infarction, coronary artery by-pass grafting, percutaneous coronary intervention, typical chest pain, present smoker, hypertension, hyperlipedemia, diabetes, peripheral vascular disease, positive family history). Expert interpretation was used as gold standard. Receiver operating characteristic (ROC) curves were calculated and the areas under the ROC curves for the networks trained with and without clinical data were compared for the diagnosis of myocardial infarction and ischaemia.
Results: There was no statistically significant difference in ROC area for the diagnosis of myocardial infarction between the neural network trained with the combination of clinical and image data (95.2%) and with image data alone (96.5%). For the diagnosis of ischaemia there was no statistically significant difference in ROC area between the neural network trained with the combination of clinical and image data (83.2.3%) and with image data alone (84.4%).
Conclusion: Neural network interpretation of MPS is not improved when clinical data is added to perfusion and functional data. One reason for this could be that experts base their interpretations of MPS mainly on the images and to a lesser degree on clinical data.
LU TP 09-07